Using Deep Networks and Transfer Learning to Address Disinformation

Numa Dhamani, Paul Azunre, Jeffrey L. Gleason, Craig Corcoran, Garrett Honke, Steve Kramer, Jonathon MorganarXiv 2019 · arXiv

TL;DR

This paper introduces a character-level Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) ensemble model for detecting disinformation across various online platforms. By leveraging transfer learning, the model effectively adapts to new tasks with limited labeled data, demonstrating strong performance in identifying spam, review bombing, and political sentiment.

What Problem Does This Paper Address?

The proliferation of online disinformation poses a significant threat to public discourse and democratic processes. Traditional methods for combating this often rely on platform-specific metadata or extensive labeled datasets, which are not always available or consistent across the diverse and fragmented online ecosystem. This paper tackles the challenge of developing a robust, content-based approach that can identify malicious communication across different platforms, even with scarce labeled data.

Methods

The core of the proposed solution is a deep learning architecture combining a character-level CNN and a bidirectional LSTM. The character-level CNN is chosen over word-level models to better handle the nuances of social media language, including misspellings, slang, and emoticons, without being constrained by fixed vocabularies. This network encodes individual sentences into feature vectors. These sentence encodings are then fed into a bidirectional LSTM, which classifies the entire document. A key aspect of their approach is transfer learning: the model is initially trained on a large, general dataset and then fine-tuned on smaller, task-specific datasets. This allows it to adapt to new disinformation detection tasks with significantly less labeled data.

Datasets

The authors utilized several datasets for training and evaluation:

  • Enron email dataset: Used for initial training of the spam email classifier.
  • 419 spam fraud corpus: Another dataset used for initial spam classification training.
  • NASA Jet Propulsion Laboratory (JPL) email abuse dataset: Also used for initial spam classification training.
  • Howard Schultz political sentiment dataset: A custom dataset of approximately 6,000 supportive posts, 20,000 negative/harmful posts, and 40,000 neutral posts, sampled for political sentiment analysis.
  • On-topic vs. off-topic movie reviews: A dataset collected from a popular crowd-sourced review website for review bombing detection.

Key Insights and Ideas

  • Character-level CNNs are powerful for noisy social media text: The ability to process text at the character level bypasses the limitations of fixed vocabularies and effectively handles misspellings, slang, and other informal language common in online discourse.
  • Transfer learning is crucial for disinformation detection: The scarcity of labeled data in disinformation contexts makes transfer learning an essential technique. The model can learn general language representations and then adapt them to specific tasks with minimal supervision.
  • A unified approach for diverse disinformation types: The proposed CNN-LSTM architecture demonstrates versatility, successfully addressing distinct disinformation challenges like spam, review manipulation, and political sentiment analysis.
  • Latent representation captures discourse nuance: The learned latent space provides a richer representation of discourse compared to simpler methods like bag-of-words, making it more suitable for complex analysis and exploratory tasks.

Key Results

The model achieved strong performance across various tasks:

  • Spam Email Classification: A binary test accuracy of 96.14% was achieved on a balanced dataset of 7000 samples per class. After transfer learning to a multi-class classifier with 8 classes, the accuracy was 93.25%, with a precision of 0.96, recall of 0.49, and F1 score of 0.65.
  • Review Bombing Detection: Achieved a binary test accuracy of 99.5% on on-topic vs. off-topic movie reviews, indicating high effectiveness in detecting inauthentic review behaviors.
  • Political Sentiment Analysis: The model achieved a binary test accuracy of 88.10% on political sentiment classification. Notably, the accuracy for classifying negative/harmful posts was 100%, with a precision of 0.68, recall of 0.98, and F1 score of 0.81.

Limitations and Future Work

Limitations:

  • The authors acknowledge that simpler approaches like bag-of-words might achieve comparable performance on some supervised tasks. However, they argue their method captures richer discourse nuances.
  • No hyper-parameter tuning was performed for the political sentiment task, suggesting potential for further metric improvement.

Future Work:

  • The authors are exploring an ongoing effort to examine the dynamics of language usage across communities and social platforms. This includes measuring language similarity, tracking changes over time, and identifying sources of language diffusion.
  • They envision their model as a piece of a broader approach to combat disinformation, aiming to develop more comprehensive solutions.

Summary updated

BibTeX
@misc{dhamani2019using,
  author = {Numa Dhamani and Paul Azunre and Jeffrey L. Gleason and Craig Corcoran and Garrett Honke and Steve Kramer and Jonathon Morgan},
  title = {{Using Deep Networks and Transfer Learning to Address Disinformation}},
  year = {2019},
  eprint = {1905.10412},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1905.10412},
}
Figure